Brain Tumor Segmentation Using Parallel U-Net and Multi-View Fusion

Glioma is one of the most significant causes of cancer, and early diagnosis is crucial for treatment and increasing patient survival. Magnetic Resonance Imaging (MRI) is the primary method used for detecting brain tumors. Segmentation of medical images using deep learning has been established as an accurate and reliable tool for detecting glioma. Challenges in tumor segmentation using Deep Convolutional Neural Networks (DCNNs) are due to variations in intensity and diverse appearance of tumors. In this research, we implemented two different strategies for segmenting brain tumors. The first strategy uses a parallel hybrid network based on the classic U-Net architecture. One path incorporates $3 \times 3$ kernels to detect boundaries and fine details in smaller tumors, while the other path uses $5 \times 5$ kernels to capture important features in larger tumors. The second method utilizes a multi-view fusion approach. By employing this approach, we utilize contextual information from 3D input images while using a 2D network. It is worth emphasizing that 2D techniques have been more successful at dealing with clinical problems than 3D methods. The results demonstrate that our method performs competitively with existing models. Our proposed network was assessed using the BraTS 2018 dataset, achieving average Dice coefficients of 0.814 for the tumor core (TC), 0.887 for the whole tumor (WT), and 0.825 for the enhancing tumor (ET) segments.